[AMD] Implement RepOrder for AMD MMA layouts (#5126)

Implement RepOrder methods for MFMA and WMMA layouts. Both layouts have row major rep layout. Also, isTranspose flag in MFMA layout does not affect RepOrder, meaning RepOrder is row major in both cases. Co-authored-by: Ognjen Plavsic <[email protected]>
triton-lang · Nov 18, 2024 · 66e8629 · 66e8629
1 parent 0bd30a2
commit 66e8629
Show file tree

Hide file tree

Showing 2 changed files with 23 additions and 3 deletions.
diff --git a/include/triton/Dialect/TritonGPU/IR/TritonGPUAttrDefs.td b/include/triton/Dialect/TritonGPU/IR/TritonGPUAttrDefs.td
@@ -777,6 +777,11 @@ def MmaEncodingTrait : AttrInterface<"MmaEncodingTrait"> {
                     "getSizePerThreadForOperand",
                     (ins "int":$opIdx,
                          "int":$kWidth)>,
+
+    InterfaceMethod<"Get the order of reps (tiles of this layout that tile the whole tensor). The fastest-changing axis first",
+                    "SmallVector<unsigned>",
+                    "getRepOrderForOperand",
+                    (ins "int":$opIdx)>,
   ];
 }
 

diff --git a/lib/Dialect/TritonGPU/IR/Dialect.cpp b/lib/Dialect/TritonGPU/IR/Dialect.cpp
@@ -1658,7 +1658,14 @@ AMDMfmaEncodingAttr::getInstrShapeForOperand(int kWidth, int opIdx) const {
 }
 
 SmallVector<unsigned> AMDMfmaEncodingAttr::getRepOrder() const {
-  llvm::report_fatal_error("NYI. AMDMfmaEncodingAttr::getRepOrder");
+  auto rank = getWarpsPerCTA().size();
+  return getMatrixOrder(rank, /*rowMajor*/ true);
+}
+
+SmallVector<unsigned>
+AMDMfmaEncodingAttr::getRepOrderForOperand(int opIdx) const {
+  auto rank = getWarpsPerCTA().size();
+  return getOrderForDotOperand(opIdx, rank, /*kMajor*/ true);
 }
 
 SmallVector<int64_t>
@@ -1745,8 +1752,16 @@ AMDWmmaEncodingAttr::getShapePerCTATile(ArrayRef<int64_t> tensorShape) const {
   return shapePerCTATile;
 }
 SmallVector<unsigned> AMDWmmaEncodingAttr::getRepOrder() const {
-  llvm::report_fatal_error("NYI. AMDWmmaEncodingAttr::getRepOrder");
+  auto rank = getWarpsPerCTA().size();
+  return getMatrixOrder(rank, /*rowMajor*/ true);
 }
+
+SmallVector<unsigned>
+AMDWmmaEncodingAttr::getRepOrderForOperand(int opIdx) const {
+  auto rank = getWarpsPerCTA().size();
+  return getOrderForDotOperand(opIdx, rank, /*kMajor*/ true);
+}
+
 SmallVector<unsigned> AMDWmmaEncodingAttr::getCTAsPerCGA() const {
   return SmallVector<unsigned>(getCTALayout().getCTAsPerCGA());
 }
@@ -2016,7 +2031,7 @@ NvidiaMmaEncodingAttr::getSizePerThreadForOperand(int kWidth, int opIdx) const {
 // DotOperand Encoding
 //===----------------------------------------------------------------------===//
 SmallVector<unsigned> DotOperandEncodingAttr::getRepOrder() const {
-  if (auto mma = mlir::dyn_cast<NvidiaMmaEncodingAttr>(getParent())) {
+  if (auto mma = mlir::dyn_cast<MmaEncodingTrait>(getParent())) {
     return mma.getRepOrderForOperand(getOpIdx());
   }
   llvm::report_fatal_error(